Paste this rule into Claude Code, and it will start thinking about how it feels when reading code

Khairallah AL-Awady posted a tweet telling vibecoders to paste a set of rules into Claude Code or Grok Bot. It wasn't long—just two rules.

First: Before any change, check it against all three—UX, DX, and AX. If it makes one win and another suffer, stop and tell me; don't just ship it.

Second: Keep files small, logic flat, and naming honest. AX is the reader you forgot about—and the one reading your code the most right now.

UX and DX are nothing new. User experience, developer experience—product managers and engineers have been arguing about them for over a decade. What's new is AX. Agent experience. Your code isn't just for people to read; it's also for models to read. A model can't understand your abbreviations, can't guess whether your variable named data holds a list of users or a cache dictionary, but it won't raise its hand to ask. It will just keep writing based on what it thinks it means.

A commenter added a sharp follow-up. ChronoNinja said that on GitHub, the share of pull requests initiated by AI agents was 17% in March and had exceeded 40% by September. His exact words were: The first draft isn't yours anymore, but the last word still is.

Put that data point next to 'honest naming' and the point is clear. If a variable called temp actually stores a user's balance, a human might spot it after scrolling two screens of code. An agent won't. It will take the name at face value, then build three layers of logic on top of it. By the time you see it in code review, it's already too late.

Another reply was more specific. A commenter said passing tests are the metric he trusts least. He compared one tab against the design mockup and measured 981 issues. Tests read code; nobody reads the screen. ChronoNinja replied: 981 is brutal. Tests read code, users read the screen, and no merge rate can reflect that gap.

That's the value of the 'stop and tell me' rule. Agents are very good at finishing things, but not at judging whether the thing should be done. Ethan Halfhide's reply puts it bluntly: he has killed many agent 'wins' because those changes let the agent get through, but made him slower as the operator. If the agent ships and he drowns in cleanup, that's not shipping.

One more reply deserves to be called out separately. Sebastian Buzdugan said the first draft can be written by an agent, but retry logic cannot. His exact words were: I won't let it touch anything like 'will retries charge the customer twice?'

So what this rule is really saying isn't 'make the agent write better.' It's saying that when agents start writing a large part of your codebase, your coding conventions need to be examined from a different angle. File size, logic depth, variable naming—these used to be style issues; now they're interface issues. The agent is the biggest caller of that interface.

Paste it in and try it. You don't need to change any tool configuration—just throw those two sentences into Claude Code or your agent's rules file. Watch when it starts stopping to ask you instead of quietly opening the PR.

That pause is where AX starts to work.